planning-with-files

Pass

Audited by Gen Agent Trust Hub on Mar 9, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill establishes a vulnerability surface for indirect prompt injection through its core file-based workflow. \n
  • Ingestion points: Untrusted data enters the agent's context through research findings stored in notes.md and tool/system errors logged in task_plan.md (as outlined in SKILL.md and examples.md). \n
  • Boundary markers: The provided templates for persistent files lack delimiters or instructions for the agent to distinguish between its own structured planning and untrusted external content. \n
  • Capability inventory: The skill requires continuous use of read, write, and edit tool calls on the local filesystem to maintain its 'working memory'. \n
  • Sanitization: No instructions are provided for sanitizing or escaping data gathered from external sources before it is committed to these persistent files. \n
  • Vulnerability: Principle 2 ('Attention Manipulation Through Repetition') in reference.md specifically instructs the agent to re-read the plan file before every decision to 'refresh goals', which brings any malicious instructions embedded in the logs directly into the agent's active prompt window.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 9, 2026, 10:12 PM